Improved YOLO architectures for efficient and accurate object detection in aerial images

Aerial object detection plays a vital role in autonomous Unmanned Aerial Vehicle (UAV) applications such as surveillance, disaster response, and precision agriculture. However, aerial imagery presents unique challenges, including wide-scale variation, dense object distributions, and cluttered backgrounds. While the YOLO family of models has demonstrated excellent real-time performance, lightweight variants such as YOLOv11s still exhibit limited accuracy on aerial datasets. This thesis proposes a series of architectural improvements to YOLOv11s to enhance its performance on aerial images without compromising efficiency. The proposed modifications include additional prediction heads (P2–P5), attention mechanisms (CBAM, GAM), alternative feature-fusion networks (BiFPN), lightweight backbone substitutions (Ghost, EfficientNet), and optimized loss functions (WIoU, Focal ECIoU). Extensive experiments were conducted on two benchmark datasets, VisDrone and DOTA, to evaluate the trade-offs among detection accuracy, inference speed, and model size. On VisDrone, the YOLOv11s + P2–P5 + Rep variant achieved the highest accuracy improvement, raising mAP50 from 0.468 to 0.494 (+5.6%) with real-time performance (196 FPS) and compact size (23 MB). Similarly, the +P2–P5 + BiFPN and +P2– P5 variants reached mAP50 = 0.493 and 0.491, respectively, both operating above 250 FPS. On DOTA, the +P2 and +P2–P5 + Rep configurations improved accuracy to 0.561 (+1.8%) and 0.557 (+1.1%), respectively, with significant speed gains (up to 116 FPS). In contrast, the EfficientNet backbone reduced model size by 52% but caused a sharp accuracy drop, confirming the importance of balanced architectural design. Overall, this work provides a systematic analysis of YOLOv11s variants tailored for aerial object detection and highlights the trade-offs between accuracy, computational efficiency, and model compactness. The findings demonstrate that modular and lightweight architectural adaptations can substantially enhance real-time UAV perception systems without increasing model complexity.